此内容尚不支持你的语言。
The Foundation is one person and a team of AI agents. How We Work describes that arrangement. This page shows it: every agent definition in the Foundation’s knowledge base, what each one is for, and the exact set of tools each one is permitted to use.
The point is not the count. The point is that the count is checkable, that each agent’s authority is written down before it runs, and that a reader outside the Foundation can see the boundaries rather than take our word for them. A Constrained Organization that will not publish its own constraints is asking for the trust it claims to make unnecessary.
| Field | Published | Why |
|---|---|---|
| Name | Yes | The identifier used when work is delegated to the agent. |
| Description | Yes, verbatim | The selection criterion. This is the text that decides which agent receives a task. |
| Domain | Yes | The directory the definition lives in, which is how the roster is namespaced. |
| Tool surface | Yes | The operative constraint. It is what the agent may touch. |
| Agent body (system prompt) | No | Working instructions, review checklists and internal file paths. Publishing them exports the Foundation’s operating detail without telling a reader anything about what the agent is allowed to do. |
| Model assignment | No | An execution detail that changes with vendor releases and carries no governance meaning. |
The export script enforces this. It reads only the YAML frontmatter of each definition and never loads the body, so the boundary is a property of the pipeline rather than a promise about editorial care. The script is scripts/content/generate-agent-directory.mjs in the website repository, and it fails loudly if an agent appears without a name, without a description, or in a domain that has no published label.
CARE specifies five constraint dimensions: Financial, Operational, Temporal, Data Access, and Communication. The tool surface below is the machine-enforced part of two of them. An agent whose tools are Read, Glob and Grep can search the knowledge base and read what it finds. It cannot edit a file, run a command, open a network connection, or delegate to another agent. That is not a policy the agent is asked to respect. It is a list the runtime enforces before the agent’s first turn.
Three postures appear in the roster:
Read only. The agent observes and reports. Every reviewer, verifier, critic and auditor sits here, which is deliberate: the agents that judge the Foundation’s work are the ones least able to change it. security-reviewer can flag a problem in a constitutional draft and cannot fix it. claims-verifier can mark a citation fabricated and cannot delete it.
Read and write. The agent may edit files or run commands. Five definitions carry this authority, and they are the implementation and release roles rather than the judgment roles.
Inherits session tools. The definition declares no narrowing, so the agent operates with whatever the session already granted. This is the weakest of the three and it is the honest label for it. These agents are constrained by the rules and hooks that govern the session, not by a declaration of their own.
Tool surface bounds the Execution Plane. It says nothing about authority, which is the Trust Plane’s business and stays with the human:
/todos stops for human approval, and no agent can advance past that gate.Enforcement is layered rather than trusted to any single mechanism. Rules are re-injected into the session on every turn so they survive context compression. Hooks run as deterministic scripts and block a violating operation with a non-zero exit rather than asking the model to reconsider. The phase commands hold the approval gates. This is the five-layer CO architecture applied to the Foundation’s own operations.
One agent per published specification, plus the constitution and the governance-layer positioning.
care-expert Use this agent for questions about the CARE framework (Collaborative Autonomous Reflective Enterprise), the Dual Plane Model, Mirror Thesis, Human-on-the-Loop governance, the six human competency categories, or enterprise AI governance philosophy. Expert in CARE's governance architecture and its relationship to EATP and COC.
Tool surface: Read Glob Grep
care-implementation-expert CARE implementation expert. Use for Trust Plane, constraint envelopes, posture progression, gradient.
Tool surface: Read Glob Grep
co-domain-expert CO domain application expert. Use for adapting CO to Compliance, Finance, Operations, or new verticals.
Tool surface: Read Glob Grep
co-expert Use this agent for questions about Cognitive Orchestration (CO), the domain-agnostic base methodology for structuring human-AI collaboration. Expert in CO's eight first principles, five-layer architecture, and how CO relates to domain applications like COC (Codegen), COR (Research), COE (Education), COG (Governance), COL-F (Finance education), and CO for Compliance.
Tool surface: Read Glob Grep
coc-expert Use this agent for questions about Cognitive Orchestration for Codegen (COC), the five-layer architecture for AI-assisted development, vibe coding critique, anti-amnesia patterns, institutional knowledge engineering, or Human-on-the-Loop development methodology. Expert in COC as the CARE framework applied to software development.
Tool surface: Read Glob Grep
coe-expert Use this agent for questions about CO for Education (COE), the Mirror Thesis applied to university assessment, the COE Spectrum (four levels of student freedom), anti-gaming architecture, EATP for academic integrity, or process-based assessment of student-AI collaboration. Expert in COE as a domain application of CO.
Tool surface: Read Glob Grep
cog-expert Use this agent for questions about CO for Governance (COG), the application of Cognitive Orchestration to research, standards development, constitutional design, academic publication, and strategic positioning. Expert in how this repository IS a COG implementation — the Foundation running on its own methodology.
Tool surface: Read Glob Grep
constitution-expert Use this agent for questions about the Terrene Foundation constitution, Singapore CLG governance, entrenched provisions, phased governance, board composition, membership criteria, anti-capture mechanisms, IP protection, institutional partners, or ACRA filing requirements. Expert in the constitution's design rationale and companion documentation.
Tool surface: Read Glob Grep
eatp-expert Use this agent for questions about the Enterprise Agent Trust Protocol (EATP), trust lineage, agent attestation, delegation chains, verification gradient, trust postures, cascade revocation, or governance integration. Expert in EATP specification, trust operations, and implementation patterns.
Tool surface: Read Glob Grep
governance-layer-expert Use this agent for questions about the Governance Layer Thesis, Two-Era framework (Execution vs Institutionalization), five propositions, positioning relative to execution tools (Claude Code, Cursor, Windsurf), government engagement framing, or publication strategy for CO/CARE/EATP.
Tool surface: Read Glob Grep
pact-expert PACT expert. Use for D/T/R grammar, operating envelopes, knowledge clearance, or verification gradient.
Tool surface: Read Glob Grep
The COR (Cognitive Orchestration for Research) roster: discovery, verification, adversarial review, voice.
argument-critic Adversarial argument critic. Use for hostile review of claims, evidence quality, or logical structure.
Tool surface: Read Glob Grep
claims-verifier Claims verifier. Use to check claims against cited sources (VERIFIED/OVERCLAIMED/FABRICATED).
Tool surface: Read Glob Grep WebSearch WebFetch
cross-reference-auditor Cross-reference auditor. Use for citation integrity and scope boundaries across companion papers.
Tool surface: Read Glob Grep
field-expert Field expert. Use for domain knowledge, historical context, debates, or connections to your argument.
Tool surface: Read Glob Grep WebSearch WebFetch
literature-researcher Literature researcher. Use for systematic discovery, paper assessment, or citation verification.
Tool surface: Read Glob Grep WebSearch WebFetch
voice-discriminator Voice discriminator. Use to compare AI prose against the author's authentic voice baselines (CT 2026 / CS 2015).
Tool surface: not narrowed; the agent inherits whatever the session grants
writing-partner Writing partner. Use for paragraph-level co-writing with margin notes and deliberation records.
Tool surface: Read Glob Grep
The COG policy roster for Singapore regulatory and institutional work.
grant-proposal-drafter SG grant proposal specialist. Use for NRF/IMDA/AISG/EDBI/SkillsFuture proposals with partner-led applicant routing and Terrene boundary intact.
Tool surface: not narrowed; the agent inherits whatever the session grants
policy-analyst SG policy/legislation analyst. Use to parse regulatory documents and frameworks, extract structured requirements, and identify policy gaps.
Tool surface: not narrowed; the agent inherits whatever the session grants
public-translator Policy-to-public communication specialist. Use to translate CON positions into citizen-facing materials for T7 without caricature or condescension.
Tool surface: not narrowed; the agent inherits whatever the session grants
sector-impact-assessor Per-vertical AI impact analyst. Use for finance, health, manufacturing, logistics, retail, public sector impact assessments grounded in A-tier sources.
Tool surface: not narrowed; the agent inherits whatever the session grants
sg-regulatory-mapper SG AI regulatory landscape mapper. Use to produce structured maps of NAIS 2.0, IMDA/PDPC/MAS/MOH/MOE/MOM frameworks and their interactions.
Tool surface: not narrowed; the agent inherits whatever the session grants
stakeholder-cartographer Politics-aware SG institutional mapper. Use for named stakeholder maps, champions, sensitivities, prior commitments, and coalition geometry.
Tool surface: not narrowed; the agent inherits whatever the session grants
whitepaper-architect CON whitepaper/portfolio architect. Use for document structure, tier composition, per-paper briefs, register discipline, and portfolio coherence.
Tool surface: not narrowed; the agent inherits whatever the session grants
The COE roster: assessment design, rubrics, learning-outcome mapping, pedagogical review.
ai-detection-strategist GenAI assessment strategy specialist. Develops strategies for assessment integrity in the GenAI era through design rather than detection.
Tool surface: not narrowed; the agent inherits whatever the session grants
assessment-designer Assessment task design specialist. Designs assessments aligned to learning outcomes that test genuine understanding and resist trivial AI completion.
Tool surface: not narrowed; the agent inherits whatever the session grants
learning-outcomes-mapper Learning outcome mapping specialist. Maps learning outcomes across courses, programs, and accreditation standards to ensure coverage and alignment.
Tool surface: not narrowed; the agent inherits whatever the session grants
pedagogical-reviewer Pedagogical quality reviewer. Reviews assessments and course materials for alignment, fairness, accessibility, and pedagogical soundness. Never says "this is fine."
Tool surface: not narrowed; the agent inherits whatever the session grants
rubric-builder Rubric construction specialist. Builds transparent, criterion-referenced rubrics with explicit criteria, performance levels, and descriptors.
Tool surface: not narrowed; the agent inherits whatever the session grants
Task tracking, GitHub issue management, and release workflow.
gh-manager GitHub issue and project management across terrene-foundation org.
Tool surface: not narrowed; the agent inherits whatever the session grants
git-release-specialist Git workflow specialist for branch management, PR workflows, and version control of the knowledge base. Use before commits or when preparing releases.
Tool surface: Read Write Edit Bash Grep Glob Task
todo-manager Project task tracking for ecosystem-level work items.
Tool surface: not narrowed; the agent inherits whatever the session grants
Review, analysis, coordination and publication. These agents work across every workspace.
care-platform-architect CARE Platform architect. Use for multi-workspace design, knowledge package portability, or open vs commercial.
Tool surface: Read Glob Grep
claude-code-architect CC artifact architect. Use for auditing, designing, or improving agents, skills, rules, commands, hooks.
Tool surface: Read Write Edit Grep Glob Bash Task
content-flow-coordinator Website content sourcing from foundation, publications, atelier, loom.
Tool surface: not narrowed; the agent inherits whatever the session grants
deep-analyst Deep analysis for failure points, risks, and requirements. Use for complex governance decisions, constitutional analysis, or strategy evaluation.
Tool surface: Read Grep Glob Task
ecosystem-coordinator Cross-repo orchestration, content flow, dependency tracking.
Tool surface: not narrowed; the agent inherits whatever the session grants
gold-standards-validator Knowledge base compliance validator. Checks documents against Terrene Foundation naming conventions, licensing accuracy, terminology standards, and cross-reference integrity.
Tool surface: Read Glob Grep LS
intermediate-reviewer Review specialist for document quality, consistency, and cross-reference accuracy. Use after significant document changes.
Tool surface: Read Grep Glob Task
open-source-strategist Licensing and open/proprietary boundary strategy. Use for licensing, SDK architecture, community, or positioning.
Tool surface: Read Glob Grep WebSearch WebFetch
oss-coordinator OSS stack health, COC sync status, release coordination.
Tool surface: not narrowed; the agent inherits whatever the session grants
publication-expert Use this agent for academic publication preparation, venue selection, related work analysis, citation management, abstract writing, formatting for specific venues (arXiv, SSRN, AIES, AI & Society), and submission packaging. Expert in academic writing conventions and multi-venue submission strategy.
Tool surface: Read Glob Grep WebSearch WebFetch
publication-preflight Pre-submission deep validator. Use to catch fabricated refs, overclaims, factual errors before submission.
Tool surface: not narrowed; the agent inherits whatever the session grants
requirements-analyst Requirements analysis for systematic breakdown and decision records. Use when starting complex governance initiatives or specification work.
Tool surface: Read Write Edit Grep Glob Task
security-reviewer Security and sensitivity reviewer for governance documents. Use before commits involving constitutional, partnership, or strategy content.
Tool surface: Read Grep Glob
Four, stated plainly.
The tool surface is one constraint dimension made machine-readable. Financial, Temporal and Communication constraints are governed by session policy and by the human running the session, and they are not declared per agent. A reader should not read this page as a complete constraint envelope.
Nineteen of the forty-six definitions declare no tool narrowing at all. That is a real gap in the roster rather than a design choice, and publishing it is how it gets closed.
The roster is a snapshot. It is regenerated from the knowledge base when the agent set changes, and the file it renders records the run that produced it. A directory published this way can be stale; it cannot be wrong about what it saw.
Agents defined inside the Foundation’s submodules and in external repositories are not shown. This page covers the root roster, which is the set that operates across the whole knowledge base.